Why only some patients respond to immunotherapy. Immune checkpoint inhibitors targeting the PD-1/PD-L1 pathway have transformed lung cancer treatment, but durable responses occur in only 20 to 30 percent of patients. The inability to predict which patients will respond creates a clinical dilemma: patients who will not benefit may be exposed to serious immune-related toxicities, while responders may be identified too late for optimal treatment sequencing.
The sampling bias problem with PD-L1 testing. The current clinical standard for selecting immunotherapy candidates is PD-L1 immunohistochemistry performed on needle biopsy samples. However, approximately 30 percent of lung cancer patients show spatial variability in PD-L1 expression across the tumor, meaning a single biopsy sample may not represent the immune landscape of the entire lesion. This spatial heterogeneity directly leads to patient misclassification and incorrect treatment decisions.
Why static imaging is insufficient. Standard CT scans capture a single moment in time and cannot reflect the dynamic remodeling of the tumor immune microenvironment that occurs during treatment or between scans. Rapid tumor growth under immune pressure is associated with PD-L1 upregulation through HIF-1a signaling and stromal fibrosis that physically excludes CD8+ T cells. A model that analyzes only one time point misses these immune evasion dynamics entirely.
The opportunity for longitudinal radiomics. Sequential CT scans provide a non-invasive window into how tumors change over time. By quantifying tumor growth kinetics from longitudinal imaging and combining this with deep learning-based segmentation, it becomes possible to build a surrogate marker for the immunosuppressive tumor microenvironment without requiring repeated biopsies.
The core innovation: MK-UNet. MK-UNet is a 3D U-Net-based segmentation model that integrates clinical metadata directly into its architecture. Unlike standard deep learning models that process imaging data alone, MK-UNet encodes six clinical and tumor biomarkers: age, gender, smoking history, pathological grade, tumor stage, and log volume change rate (LVCR). These are encoded into a six-dimensional vector and injected at the bottleneck layer of the encoder, providing the network with biological context during feature learning.
Multi-modal preprocessing to enhance input quality. Raw CT volumes are converted from single-channel to four-channel inputs using three complementary algorithms. The first channel preserves original Hounsfield Unit values. The second applies Canny edge detection to accentuate tumor boundaries. The third uses a 3D Gaussian filter for noise reduction. The fourth performs contrast optimization by linearly scaling pixel intensities to the full 256 gray-level range. An additional adaptive windowing filter masks non-tumor voxels by computing an intensity threshold based on the mean and standard deviation of annotated tumor regions, reducing background noise before the network processes the volume.
Attention gates on skip connections. At each skip connection between encoder and decoder, an attention gate (AG) dynamically weights encoder features using contextual information from coarser decoder layers. A learnable convolution integrates encoder and decoder signals, and a sigmoid activation produces a spatial attention map that emphasizes tumor boundaries while suppressing irrelevant background. This mechanism allows the decoder to focus on the most diagnostically relevant regions as it reconstructs the segmentation mask.
Training configuration and dataset. The dataset consisted of 420 lung CT scans from 420 patients at the Fourth Hospital of Hebei Medical University, all with pathologically confirmed malignant pulmonary nodules and at least two CT examinations. Scans were split into 300 training, 60 validation, and 60 test cases. Tumor ROIs were manually delineated by two board-certified radiologists with disagreements adjudicated by a senior radiologist with over 15 years of experience. Training used the Adam optimizer at a learning rate of 0.0001, a batch size of 2, and 150 epochs on an NVIDIA RTX 4090.
What LVCR measures. The Log Volume Change Rate (LVCR) is a quantitative measure of tumor proliferative aggressiveness derived from sequential CT scans. It captures the weighted rate of logarithmic volume change over time across multiple measurements, with higher LVCR values indicating faster-growing tumors. Tumors with high LVCR above 0.15 per day are typically associated with hypoxic microenvironments that upregulate PD-L1 and promote stromal fibrosis that impedes T-cell infiltration.
The immune-aware attention gate (IAAG) mechanism. The IAAG extends the standard attention gate by incorporating LVCR alongside encoder and decoder feature maps in the attention weight computation. The gate computes a combined representation of imaging patterns and growth kinetics, then applies a sigmoid activation to produce spatial weights. In high-LVCR regions, IAAG prioritizes voxels with spiculated margins and necrotic cores, which are radiological signatures of immune-evasive tumor biology. In slow-growing lesions, it suppresses calcifications and benign hyperattenuating artifacts.
Performance benefit of IAAG for aggressive tumors. For the subset of 25 high-LVCR tumors in the test set, MK-UNet with IAAG achieved a Dice coefficient of 0.79 plus or minus 0.04 and HD95 of 8.2 plus or minus 1.3 mm. The standard 3D U-Net scored 0.71 Dice and 12.1 mm HD95 on this same subset, while Attention U-Net achieved 0.75 Dice and 10.5 mm HD95. The IAAG-driven improvement was largest on this challenging subset, confirming that the mechanism specifically addresses the segmentation failure modes associated with rapidly evolving immune-suppressive tumors.
Biological validation of the IAAG approach. Post-hoc radiomic analysis of MK-UNet segmentation masks revealed that tumors with irregular margins (sphericity below 0.6) exhibited significantly lower CD8+ T-cell density (r=0.41, p=0.007) and higher PD-L1 expression (r=0.38, p=0.01) compared to spherical tumors. Independent corroboration from TCGA-LUAD data showed that CT-based spiculation length positively correlated with PD-L1 positivity (r=0.38, p=0.002), validating the biological plausibility of using morphological imaging features to infer immune microenvironment status.
Overall segmentation results. On the 60-case test set, MK-UNet achieved a Dice coefficient of 0.7728 plus or minus 0.03 and IoU of 0.6471, surpassing all baseline models. The 3D U-Net reached 0.7322 Dice and 0.6223 IoU. 3D ResNet achieved 0.7231 Dice. Attention U-Net reached 0.7527 Dice and 0.6302 IoU. The improvement over Attention U-Net demonstrates that integrating clinical metadata provides gains beyond what imaging-only attention mechanisms can achieve.
Boundary accuracy measured by Hausdorff distance. HD95, which measures the maximum deviation of the predicted boundary from the true boundary at the 95th percentile, was 9.8 plus or minus 1.5 mm for MK-UNet, compared to 12.4 plus or minus 2.1 mm for 3D U-Net and 11.2 plus or minus 1.8 mm for Attention U-Net. This 21 percent reduction in boundary error is clinically significant for radiotherapy planning, where precise tumor delineation directly determines dose coverage to the target and sparing of surrounding healthy tissue.
Ablation study confirms each component's contribution. The stepwise ablation isolated the contribution of each innovation. Starting from a baseline 3D U-Net at Dice 0.7322 and HD95 12.4 mm, adding multi-modal preprocessing improved Dice to 0.7516 and HD95 to 11.5 mm. Further adding medical knowledge fusion (clinical metadata) raised Dice to 0.7634 and HD95 to 10.3 mm. Adding lesion-aware attention achieved the final Dice of 0.7728 and HD95 of 9.8 mm. Each step produced consistent gains, confirming a synergistic relationship among the three components.
Computational efficiency for clinical deployment. MK-UNet contains 28.5 million parameters compared to 27.9 million for standard 3D U-Net, demonstrating that the added clinical knowledge fusion and attention mechanisms required minimal additional parameters. Inference time is 15 seconds per scan at 512 by 512 by 40 voxels, compatible with real-time clinical workflows. The model outputs DICOM-format binary segmentation masks, enabling direct integration with commercial radiotherapy planning systems including Eclipse, Monaco, and Pinnacle.
Defining the Immune Evasion Score (IES). Beyond segmentation, the study introduces the IES as a composite biomarker derived from the segmented tumor mask. The formula combines three weighted components: IES = 0.6 times LVCR plus 0.3 times (1 minus Sphericity) plus 0.1 times PD-L1-positive Area Ratio. The weights were determined by multi-factor Cox regression analysis based on the prognostic strength of each predictor: LVCR hazard ratio 1.8 (p less than 0.001), sphericity hazard ratio 1.3 (p=0.02), and PD-L1 positive area ratio hazard ratio 1.1 (p=0.04).
IES predicts PD-L1 expression and T-cell exclusion. Validation on the dataset showed that the IES predicted PD-L1 expression with an AUC of 0.85 (p less than 0.001) and was significantly associated with CD8+ T-cell exclusion (p less than 0.01). A spatial analysis comparing IAAG-prioritized regions with pathological PD-L1-positive niches in the tissue showed 72.3 percent concordance, demonstrating that the model's attention maps spatially align with biologically confirmed immune-suppressive zones.
IES predicts immunotherapy progression. High IES was associated with significantly faster progression on immunotherapy, with a hazard ratio of 2.3 (p=0.004). This means patients with high IES had more than double the risk of rapid disease progression compared to patients with low IES. This predictive relationship was established using survival analysis with the Schoenfeld residual test confirming that the proportional hazards assumption was satisfied.
Clinical stratification potential. The IES framework enables risk-stratified treatment planning. Patients with high IES indicating active immune evasion may benefit from combination strategies using PD-1/PD-L1 inhibitors plus chemotherapy, while patients with low IES may be candidates for immunotherapy monotherapy. Sequential IES computed from follow-up CT scans can track temporal changes in immune microenvironment status, potentially guiding decisions about treatment continuation or modification.
Addressing the clinical gap in dynamic monitoring. Existing AI-based segmentation models treat tumors as static anatomical structures and do not incorporate longitudinal growth dynamics. MK-UNet fills this gap by embedding LVCR as a temporal prior, enabling the model to adapt its feature extraction to how the tumor is evolving rather than just what it looks like at a single time point. In pseudo-progression cases where inflammatory changes mimic tumor growth, LVCR-guided filtering reduced false positives by 18 percent compared to Attention U-Net.
Implications for radiotherapy planning. MK-UNet's Dice performance of 0.7728 and HD95 of 9.8 mm falls within the inter-observer variability reported among radiologists (HD95 range 8.2 to 10.1 mm), suggesting it approaches human-level consistency for delineation tasks. Producing biologically informed segmentation masks that preserve pleural infiltrations and necrotic cores, which are often oversmoothed by standard models, directly supports more accurate dose planning for advanced adenocarcinomas.
Study limitations requiring future work. LVCR calculation requires at least two sequential CT scans, limiting applicability to patients with incidental findings on their initial scan. The training data came from a single institution with uniform imaging protocols, and external cohort characteristics were not fully reported, restricting assessment of generalizability. The model also struggles with tumors adjacent to high-attenuation structures like the chest wall where boundary ambiguity remains. Additionally, LVCR's prognostic value beyond segmentation improvement has not been independently assessed in survival studies.
Future directions. Prospective multi-center validation with diverse scanner protocols is the immediate priority. Future model iterations could incorporate anatomical priors such as organ-atlas registration to resolve chest wall boundary ambiguity, and three-dimensional texture features capturing spatial relationships between necrotic cores and viable tumor tissue could serve as additional non-invasive immune biomarkers. Establishing formal links between MK-UNet outputs and clinical endpoints including overall survival and recurrence will be essential before clinical adoption.